Visual standards: reputable neuroscience vs AI slop#

The visual bar for Kahlus is: every research figure must be generated from data, code, and provenance. Prompted SVGs are allowed only for explicitly labeled conceptual schematics.

Reputable visual patterns#

EEG / BCI / neurophysiology#

Use MNE/MOABB/Braindecode conventions:

  • raw traces are shown in real time units, with channel labels and scale clarity;

  • topomaps require actual sensor names/positions, usually via a standard montage or dataset montage;

  • PSD/time-frequency plots state frequency bands and sampling rate;

  • model overlays include actual signal, prediction, residual, and a metric;

  • benchmark plots distinguish train/val/test and avoid random-window leakage.

NeuroML / computational neuroscience#

Use NeuroML/pyNeuroML conventions:

  • morphology figures come from cell/network model files, not decorative neuron clip art;

  • voltage/current traces include units, stimulus timing, and protocol labels;

  • F-I curves and electrophysiology summaries show the simulation protocol;

  • diagrams should map directly to model components, LEMS/NeuroML files, or documented equations.

Scientific Python documentation#

Use Sphinx/MyST/PyData conventions:

  • tutorials for first-run users;

  • how-to guides for tasks;

  • explanation pages for scientific reasoning;

  • API/reference pages for code;

  • citations and exact dependency versions for reproducibility.

Slop detector#

A figure is suspicious if it has any of these:

  • no data source or provenance;

  • no units on axes;

  • synthetic-looking EEG waves with no sampling rate or channel identity;

  • topomaps without montage/channel positions;

  • unverified claims such as “ridge learns neural state” when it may be exploiting short-horizon autocorrelation;

  • too many overlapping traces for readability;

  • decorative gradients, glassmorphism, fake 3D brains, or icons replacing evidence;

  • captions that explain vibes rather than methods.

Kahlus figure rule#

Every figure must be one of:

  1. Benchmark-derived evidence: generated from stored tensors/manifests/results; includes source, split, model, and metrics.

  2. Diagnostic refit: recomputed from the same exported tensors with code matching the benchmark baseline; clearly labeled.

  3. Schematic: conceptual only; stamped SCHEMATIC - NOT BENCHMARK EVIDENCE.

Anything else does not go into mentor-facing docs.